The wells behind the lab
Six synthetic wells where the truth is known, for the lessons, and four real wells from the North Sea, for trying them out.
The teaching field
Six fictional wells, A to F, through the same layer-cake of shale, Sand A, a shale with coals, a limestone, a shale with anhydrite, Sand B and shale. Wells A, B and C are close together on the crest; D, E and F are step-outs that differ the way wells across a real field do. They're built from the rock physics in Probe Lab, so every sample carries its true facies, clay volume, porosity, water saturation and permeability. That truth is what lets every figure show where a model was right and wrong, rather than just a score.
Each well is logged with the same eight curves, named as in the FORCE 2020 data so ideas move between the two:
Six facies, all of them classes in the FORCE 2020 labels, with coal and anhydrite rare on purpose:
The field is generated once, from a fixed seed, and checked by the lab's tests every time it changes. Its simplifications are on the Assumptions page.
Real wells: FORCE 2020
The FORCE 2020 machine learning competition released logs and lithology labels for about a hundred wells in the Norwegian North Sea. Fit Lab uses four of them, about 300 m each, chosen where gamma ray, density, neutron, Pe and sonic are all present and the labels fall in the lab's six facies. Three are close together in quadrant 25; the fourth, 31/4-5, is from further north, and reads differently.
The competition's lithology codes map to the lab's facies directly: 30000 sandstone, 65030 sandstone/shale, 65000 shale, 70000 limestone, 90000 coal and 86000 anhydrite. Any other class becomes "other" and is left out of the scores. The curves used are GR, RHOB, NPHI, PEF, DTC, RDEP, CALI and DRHO, rounded, with the bit size where the competition records it.
Citation and licences
The raw competition file isn't part of the lab. scripts/prepare-force.py cuts the subset from it, and anyone with the file can repeat that.